UK Facial Age Estimation Plan Faces Renewed Accuracy and Bias Scrutiny

The UK Home Office’s planned use of facial age-estimation technology to support initial age decisions for undocumented asylum seekers is facing renewed scrutiny over its accuracy for children and its performance across demographic groups.

The government is testing the technology during 2026 with a view to introducing it at the UK border in 2027. It would be used as a supplementary tool when officials dispute or remain uncertain about the age claimed by a person arriving without reliable identity documents.

Facial age estimation differs from facial recognition. It analyzes a single image to estimate a person’s age or age range and does not attempt to identify the individual by searching a facial-image database. The Home Office says human officers will retain responsibility for the final decision.

The proposed system is intended partly to address problems with existing visual assessments. Home Office figures obtained by the Helen Bamber Foundation indicate that 755 children were incorrectly treated as adults during 2025 under the current process.

Questions have nevertheless emerged about whether an automated estimate would materially improve those decisions. Lighthouse Reports obtained a leaked internal Home Office evaluation, dated April 2025, that examined seven algorithms using approximately 2.5 million images.

The report primarily discussed an unidentified “best performing supplier.” It is unclear whether that supplier was Cognitec or another company. According to Lighthouse Reports, the system tended to overestimate the ages of minors, performed worse on female subjects, and estimated the age of girls under 18 from Sub-Saharan Africa with a mean absolute error of 4.6 years.

The Home Office subsequently awarded a three-year contract valued at £322,000 to Akhter Computers. German biometric technology developer Cognitec is listed as a subcontractor providing the facial age-estimation technology. The contract began on June 1, 2026.

Lighthouse Reports separately analyzed underlying data from Cognitec’s April 2026 submission to the National Institute of Standards and Technology’s Face Analysis Technology Evaluation program.

Its analysis found that image quality had a substantial effect on performance. On NIST’s lower-quality Border dataset, Cognitec’s system classified more than two-thirds of 16-year-olds as adults. More than one-third were classified as adults on the higher-quality Application dataset.

Lighthouse also calculated that, across the datasets it examined, Cognitec’s system estimated more than half of West African 16-year-olds to be 18 or older. Fewer than one-quarter of Eastern European 16-year-olds were placed above that threshold.

Those figures are calculations by Lighthouse Reports based on data embedded in NIST’s public report cards, rather than conclusions stated directly by NIST. NIST’s broader evaluation confirms that age-estimation performance can vary substantially according to the algorithm, image source, age, sex, and region of birth.

Cognitec said demographic performance differences affect algorithms throughout the market and can be influenced by factors including image quality, training data, and testing methodology. The company said its demographic disparities were low compared with systems of similar overall accuracy and that it was continuing to develop methods intended to reduce bias.

NIST continues to update its testing as vendors submit new systems. ID TECH recently reported that NIST added two further algorithms to its facial age-estimation evaluation, expanding the comparative data available to governments and other potential buyers.

Cognitec has also continued updating its broader facial recognition portfolio, including a next-generation matcher designed to improve performance on difficult facial poses. Performance in identity matching, however, does not necessarily predict accuracy in age estimation, which is a distinct technical task.

The Home Office emphasizes that an age estimate will form only one part of the assessment and says individuals will be treated as children when uncertainty remains. Critics argue that even a supplementary score could carry significant weight in a high-pressure border environment, particularly when other evidence is limited.

Testing will therefore need to examine not only average accuracy, but also the frequency and consequences of errors near the critical 18-year threshold and the extent to which performance observed in benchmark datasets transfers to operational border conditions.

Sources: UK Home Office, Lighthouse Reports, NIST, The Guardian

---

By the ID Tech Editorial Team